[gpfsug-discuss] slow startup of AFM flush to home
Billich Heinrich Rainer (PSI)
heiner.billich at psi.ch
Mon Oct 16 14:36:09 BST 2017
Hello Scott,
Thank you. I did set afmFlushThreadDelay = 1 and did get a much faster startup. Setting to 0 didn’t improve further. I’m not sure how much we’ll need this in production when most of the time the queue is full. But for benchmarking during setup it’s helps a lot. (we run 4.2.3-4 on RHEL7)
Kind regards,
Heiner
Scott Fadden did write:
When an AFM gateway is flushing data to the target (home) it starts flushing with a few threads (Don't remember the number) and ramps up to afmNumFlushThreads. How quickly this ramp up occurs is controlled by afmFlushThreadDealy. The default is 5 seconds. So flushing only adds threads once every 5 seconds.
This was an experimental parameter so your milage may vary.
Scott Fadden
Spectrum Scale - Technical Marketing
Phone: (503) 880-5833
sfadden at us.ibm.com
http://www.ibm.com/systems/storage/spectrum/scale
----- Original message -----
From: "Billich Heinrich Rainer (PSI)" <heiner.billich at psi.ch>
Sent by: gpfsug-discuss-bounces at spectrumscale.org
To: "gpfsug-discuss at spectrumscale.org" <gpfsug-discuss at spectrumscale.org>
Cc:
Subject: [gpfsug-discuss] AFM: Slow startup of flush from cache to home
Date: Fri, Oct 13, 2017 10:16 AM
Hello,
Running an AFM IW cache we noticed that AFM starts the flushing of data from cache to home rather slow, say at 20MB/s, and only slowly increases to several 100MB/s after a few minutes. As soon as the pending queue gets no longer filled the data rate drops, again.
I assume that this is a good behavior for WAN traffic where you don’t want to use the full bandwidth from the beginning but only if really needed. For our local setup with dedicated links I would prefer a much more aggressive behavior to get data transferred asap to home.
Am I right, does AFM implement such a ‘slow startup’, and is there a way to change this behavior? We did increase afmNumFlushThreads to 128. Currently we measure with many small files (1MB). For large files the behavior is different, we get a stable data rate from the beginning, but I did not yet try with a continuous write on the cache to see whether I see an increase after a while, too.
Thank you,
Heiner Billich
--
Paul Scherrer Institut
Science IT
Heiner Billich
WHGA 106
CH 5232 Villigen PSI
056 310 36 02
https://www.psi.ch
More information about the gpfsug-discuss
mailing list